On Burbea-Rao divergence based goodness-of-fit tests for multinomial models (Q1290938): Difference between revisions

Several statistics and families of statistics can be found in the literature on tests for goodness of fit for a vectorial parameter of proportions in a multinomial distribution. These statistics are implicitly or explicitly based on either distances, dissimilarities or divergences. The asymptotic distributions of these statistics are also known to be chi-squared (the degrees of freedom depending on the hypothesis to be tested). In this paper, the author suggests for the same purpose a generalized statistic which makes use of the family of \(R_\phi\)-divergences introduced by \textit{J. Burbea} and \textit{C.R. Rao} [IEEE Trans. Inf. Theory IT-28, 961-963 (1982; Zbl 0497.94002); ibid., 489-495 (1982; Zbl 0479.94009)]. The asymptotic distribution of this generalized statistic is first obtained under the equiprobable null hypothesis (which is also chi-squared), and an interesting comparison between particular statistics is developed via the Pitman efficiency (which is not a valuable criterion in this case for puposes of comparison), moments and exact power criteria. Then the asymptotic distributions of the statistics under composite hypotheses, as well as the asymptotic power of the test for goodness of fit, are deduced.